Essays in Economics of Artificial Intelligence

This thesis embodies three chapters on the economics and applications of artificial intelligence (AI). The first chapter explores the economic underpinnings of open-source contributions in AI by for-profit companies, focusing on large language models (LLMs). Three main findings emerge: (1) LLMs align well with the R&D portfolios of diverse technologically advanced firms, (2) models developed by large technology companies are more likely to be open-sourced, and (3) open-sourcing advanced LLMs enhances research-related activities. A theoretical framework analyzes the factors influencing a firm's decision to open-source, suggesting an inverted-U-shaped relationship between open-sourcing propensity and the firm's share of LLM-compatible applications. The second chapter addresses the debate on the moderation of toxic speech on social media and its impact on the plurality of online discourse. A new methodology is proposed and validated to measure plurality based on the semantic variance of online content, using text embeddings from computational linguistics. Applying this measure to a dataset of 10 million US Tweets, it is found that removing toxic content reduces the plurality of online discourse. Crucially, the reduction in plurality is attributed not to the toxic language itself, but to the removal of meaningful content. The third chapter proposes a novel method for estimating biases at the micro-level in contexts with multiple bilateral interactions, where individual preferences and correlated characteristics complicate analysis. The method employs Collaborative Filtering in an `honest' design to extract preferences and characteristics, separating self-induced outcomes from the constructed embeddings of interacting units.